Machine learning for dysphagia risk prediction in older adults after acute ischemic stroke

Post-stroke dysphagia (PSD) is prevalent in older adults after acute ischemic stroke (AIS), increasing risks like pneumonia and malnutrition. This study aimed to develop and validate machine learning models for early PSD prediction. In this retrospective cohort study, we utilized electronic health records from a cohort of 908 AIS patients (≥ 60 years), partitioned into training and internal test sets at a 7:3 ratio. The Least Absolute Shrinkage and Selection Operator (LASSO) regression and the Boruta algorithm were used to identify key PSD predictors. Eight ML algorithms were trained, and model performance was assessed using metrics such as the area under the curve (AUC), accuracy, specificity, recall, F1 score, Brier score, calibration curves, and decision curve analysis (DCA). Additionally, the Shapley Additive exPlanations (SHAP) method was employed for model interpretation. The incidence of PSD was 51.1% (464/908) in this cohort. Random Forest (RF) emerged as the optimal model, achieving the highest AUC of 0.873 in the testing set, slightly superior to the gradient boosting machine (0.869) and the neural network (0.867). It also exhibited competitive performance across various metrics, including accuracy (0.783), specificity (0.737), recall (0.830), F1 score (0.792), and Brier score (0.148). SHAP analysis identified hypertension, the Barthel Index, and dysarthria as the top three significant predictors of PSD among stroke patients. These findings suggest that ML algorithms, especially RF, can accurately predict PSD, allowing for early targeted interventions. Integrating ML tools into stroke care pathways could reduce dysphagia-related complications and improve resource utilization. Further prospective studies are needed to validate these results across diverse populations. Not applicable.